NVIDIA CES 2026: Rubin Architecture, Rubin Chips and Agentic AI

NVIDIA announced Rubin architecture, new chips, and agentic AI platform at CES 2026.
Key insights
Rubin architecture emphasizes liquid cooling and confidential compute
NVIDIA presented the Rubin architecture at CES 2026, specifying 100% liquid cooling using 45°C warm water and built-in confidential computing with encryption in transit, at rest and during compute. Rubin includes a Vera CPU optimized for power-constrained environments and Rubin GPUs with floating-point performance claimed at 5× Blackwell while increasing transistors by 1.6×. NVLink 6 Switch, ConnectX-9 (1.6 Tb/s), BlueField‑4 DPU and Spectrum‑X silicon photonics were listed as core platform components. Jensen Huang framed the platform as enabling a new generation of agentic AI and large-scale model training efficiency.
Rubin reduces token generation costs and expands memory via KV Cache
NVIDIA claimed Rubin delivers a throughput leap for training 100T-parameter models and that Rubin’s token-generation cost is one-tenth that of Blackwell. BlueField‑4 DPU introduces a KV Cache feature that supplies up to 16 TB of fast-access memory per GPU to mitigate GPU VRAM limits for long-context workloads. NVLink 6 and ConnectX‑9 aim to provide massive interconnect bandwidth to keep GPUs fully interconnected during large-model training.
Rubin platform includes Vera CPU, Rubin GPU and Spectrum‑X switch
Rubin platform components listed: Vera CPU with ~2× power-efficiency over prior generation and spatial multithreading; Rubin GPU with NVFP4 Tensor Core supporting dynamic precision; NVLink 6 Switch offering aggregate bandwidth asserted as multiple times global internet traffic; ConnectX‑9 NIC at 1.6 Tb/s; Spectrum‑X Ethernet switch with 512 channels using silicon photonics and co-packaged optics.
NVIDIA demos Blueprint agent framework and hybrid routing for privacy
NVIDIA showed a 'Blueprint' framework demo: a DGX Spark personal assistant that manages mail, calendar, to‑dos and robot control. The demo used a frontier model API as the 'brain', local open-source models for privacy-sensitive tasks, per‑intent routing to local or cloud models, and tool integrations (mail, calendar, camera, robot). The demo highlighted hybrid model routing and intent-based decisioning for privacy-preserving agent workflows.
Agentic AI framed as replacement interface for enterprise software
NVIDIA and the CES coverage described Agentic AI as shifting software interaction from forms and menus to conversational, goal-oriented agents. Agents were presented as multi‑modal, multi‑model, multi‑cloud and hybrid‑cloud systems able to plan, use tools, reason, and collaborate across specialist models. Demonstrations included agents composing emails, converting sketches to renderings and orchestrating camera/robot actions.
Open and inference-model advances catalyze agent ecosystems
CES commentary linked three 2025 breakpoints to the agent surge: the rise of inference/push-time scaling models (e.g., OpenAI o1/o3), the practical deployment of agentic systems that call tools and plan, and open-source model advances like DeepSeek R1. The combination was described as enabling broad agent construction across startups, enterprises and governments.